[llvm] r222760 - Reapply 222538 and update tests to explicitly request small code model

[Joerg Sonnenberger]

On Thu, Nov 27, 2014 at 11:30:40AM -0500, Rafael Espíndola wrote:
> On 27 November 2014 at 11:16, Joerg Sonnenberger
> <joerg at britannica.bec.de> wrote:
> > On Thu, Nov 27, 2014 at 11:01:54AM -0500, Rafael Espíndola wrote:
> >> This is odd at least.
> >>
> >> First,  -code-model=smal is the default, please don't specify it everywhere.
> >> Second, why the PIC change? The gnu assembler always uses a
> >> DW_EH_PE_pcrel encoding. What is the advantage of not doing that?
> >
> > I specified it explicitly to get a consistent value.
> 
> It is consistent. It is the default and it is important that the small
> model remains the default.

It is still a hidden assumption.

> >I'm not sure pcrel
> > buys anything for non-PIC, but using absolute locations is certainly
> > more consistent with what we do for CFI.
> 
> It is inconsistent with what gas does and it it looks like a
> regression. Relative encoding is more efficient and we should keep
> using it.

It adds a penalty to the reader. I don't believe it is more efficient,
especially if the size of the encoded fields is the same.

Joerg